Zhongwen Yue

To achieve high-precision identification and intelligent segmentation of explosive-induced crack structures, this study constructs a multi-lithology, multi-scale CT image dataset of blasting cracks and systematically investigates the performance differences of various deep learning-based network models for crack segmentation. Explosion loading and CT scanning experiments were first conducted on representative lithological materials, including sandstone, marble, and iron ore. Highresolution CT images of crack structures were acquired, annotated, and organized to establish a standardized explosive crack CT image database. Image quality and feature diversity were further enhanced through adaptive brightness and contrast adjustment, granular detail enhancement, and multi-scale slice reconstruction, thereby providing high-quality training samples for deep neural networks. Three semantic segmentation models-U-Net, ResU-Net, and Attention U-Net-were constructed, and a comparative evaluation of their crack recognition performance was carried out under varying lithological conditions and image brightness levels. The results demonstrate that deep learning-based approaches exhibit clear advantages over traditional threshold segmentation and edge detection methods in terms of crack continuity, noise suppression, and boundary delineation accuracy. Among the evaluated models, ResU-Net shows the most stable performance under variations in illumination and material properties, enabling accurate extraction of the true morphology and spatial distribution of explosive cracks. Its F1-score and Intersection over Union (IoU) metrics consistently outperform those of the other network architectures. Overall, the findings indicate that the ResU-Net-based deep semantic segmentation model represents an optimal solution for explosive crack CT image recognition, providing reliable technical support for elucidating explosive damage mechanisms, quantitatively characterizing crack structures, and optimizing blasting design parameters.

Over the years, the directional fracture control blasting technology using slotted charges has been widely applied in the formation control around rock tunnels. This article employs CT scanning technology and numerical simulation methods to investigate the influence of bedding angle on the blasting fracture effect in layered rock masses under different bedding strengths. The results indicate that in layered rock masses with strong bedding bond strength, the slotted charge still exhibits good directional fracture effects. The angle of strong bedding significantly affects the propagation of the explosive stress wave from the slotted charge, thereby influencing the overall damage of the specimen. Changes in the angle of the strong bedding plane have a certain impact on the expansion trajectory of the main blasting crack. As the inclination angle of the weak bedding plane increases, the influence of the weak bedding plane on the crack propagation behavior gradually increases.

The method of rock joint identification that combines machine learning with drilling parameters is effective in processing large datasets while maintaining high accuracy. However, it predominantly relies on a single parameter and does not fully account for the interactions and relative importance among different parameters. In this study, a novel two-step intelligent recognition method is proposed to classify and identify rock joints by integrating multiple fields, including drilling technology, image processing, and deep learning. Initially, a particle swarm optimization-support vector machine model is employed to classify hard and soft layer rock joints using dimensionless processed drilling parameters. For soft layer data, the gram angle field technique is introduced to convert the raw one-dimensional time series data into two-dimensional feature images. Subsequently, a convolutional neural network with residual blocks is constructed and trained on these two-dimensional feature images to classify and recognize different soft-layer rock joint images. A comparative analysis of the classification of soft layer rock joint images is then conducted using the support vector machine model, validating the effectiveness of the proposed method. The results show that the classification accuracy of this intelligent recognition approach reaches 99%, successfully identifying various joint types in the rock mass. This study establishes a foundation for the intelligent application of rock joint recognition technology and offers new perspectives for rock structure analysis and exploration in complex geological environments.